首页>
外文期刊>Journal of hydrologic engineering
>Discussion of 'Monthly Mean Streamflow Prediction Based on Bat Algorithm-Support Vector Machine' by Bing Xing, Rong Gan, Guodong Liu, Zhongfang Liu, Jing Zhang, and Yufeng Ren
【24h】
Discussion of 'Monthly Mean Streamflow Prediction Based on Bat Algorithm-Support Vector Machine' by Bing Xing, Rong Gan, Guodong Liu, Zhongfang Liu, Jing Zhang, and Yufeng Ren
The discusser wishes to thank the authors for exploring the exactness of a bat algorithm-based support vector machine (BA-SVM) in forecasting monthly streamflows of the Yangtze River in China by using previous rainfall and streamflow data. The BA-SVM results were compared with artificial neural networks (ANNs) and cross validation-based SVM (CV-SVM). The results indicated that the BA-SVM model provided better accuracy than the ANN and CV-SVM models in monthly streamflow forecasting. The discusser needs to bring up the some critical perspectives, which the authors and other potential researchers may consider: 1. Xing et al. (2015) measured the monthly mean streamflow and precipitation data for the period of 1952-2011 in Yichang Station, China. They utilized data from 1952 to 1999 for training and the remaining data for testing. A basic issue in training an ANN is abstaining from overfitting as it decreases its generalization exactness. If an excess of neurons is utilized, the network has an excess of parameters and may overfit the data. Conversely, if a couple of neurons are excessively incorporated into the network, it would not be conceivable to completely recognize the signal and fluctuation of a complex data set (Cimen and Kisi 2009). In the study, the authors used 48 years of monthly data (48 × 12 = 576 monthly values). For the optimal ANN (7,14,1) model, 112 (7 × 14+14=112) weights were utilized. The training data do not appear to be sufficient to abstain from over-fitting. The authors could use fewer hidden node numbers for the ANN models and obtain better results this way.
展开▼